Timeline of widespread 8‑bit training adoption
Determine the timeline and extent of the widespread adoption of 8‑bit numerical precision for training large-scale AI models, given routine non-disclosure of training precisions by developers.
References
Specifically, we are unsure when 8-bit training first became widespread. Developers usually do not report what precisions they use to train their models, making it difficult to assess when newly available formats were widely adopted.
— Trends in AI Supercomputers
(2504.16026 - Pilz et al., 22 Apr 2025) in Appendix: Methods — Numerical precision (Section A, Adequately representing performance gains from using lower precision units)